Glosarium AI
Kamus lengkap Kecerdasan Buatan
Policy Gradient
Direct optimization method that adjusts policy parameters by following the gradient of the expected return, enabling learning of stochastic policies without requiring an environment model.
REINFORCE Algorithm
Basic policy gradient algorithm using a Monte Carlo estimate of the gradient to update policy parameters based on fully observed episodes.
Actor-Critic Methods
Hybrid approach combining an actor that learns the policy and a critic that estimates the value function, reducing the variance of policy gradient estimates.
Advantage Function
Measure of the superiority of an action compared to the average of actions in a given state, calculated as the difference between the Q function and the V function to reduce gradient variance.
Proximal Policy Optimization (PPO)
Algorithm optimizing the policy by constraining updates to stay close to the previous policy, using a clipped objective function to ensure learning stability.
Trust Region Policy Optimization (TRPO)
Method ensuring monotonic performance improvements by optimizing the policy within a trust region defined by the KL divergence between successive policies.
Natural Policy Gradient
Variant of policy gradient using the Fisher metric to perform parameterization-invariant updates, ensuring more stable and efficient convergence.
Policy Network
Parameterized neural network that represents the policy π(a|s; θ), generating a probability distribution over actions conditioned on the current state.
Monte Carlo Policy Gradient
A gradient estimation technique that uses complete trajectories to calculate returns, providing an unbiased but high-variance estimate.
Baseline Function
A function subtracted from the return to reduce the variance of the gradient estimate without introducing bias, typically the state-value function.
Importance Sampling
A technique that allows using data collected with an old policy to update a new policy, by weighting samples according to the probability ratio of the policies.
Entropy Regularization
Adding an entropy term to the objective function to encourage exploration by penalizing overly deterministic policies, improving the robustness of learning.
Deterministic Policy Gradient
An extension of policy gradient to continuous action spaces where the policy is deterministic, particularly effective in high-dimensional environments.
Stochastic Policy
A policy represented by a probability distribution π(a|s) over actions, allowing for intrinsic exploration and is essential for policy gradient methods.
KL Divergence Constraint
A constraint that limits the Kullback-Leibler divergence between successive policies to ensure stable updates and avoid overly drastic changes in behavior.
Generalized Advantage Estimation (GAE)
An advantage estimation method that combines bias and variance through a weighted average of multi-step estimators, offering an optimal trade-off for learning.
Policy Gradient Theorem
Fundamental theorem providing an analytical expression of the gradient of the expected return with respect to the policy parameters, formulating the theoretical basis of the methods.
Return-to-Go
Sum of discounted future rewards from a given time step, used as a gradient estimator in policy gradient algorithms.